Kimi K3.1 Pro vs Claude Opus 5.5 vs GPT-6 Astra: Price, Context, and Coding Scores, Grouped by Who Ran Them

Price and specs

Who can actually use each one

Coding scores, grouped by who ran them

The same benchmark name can carry three different numbers depending on the runner, the harness, and the thinking budget, so each row names all three. 'Moonshot chart' means the K3.1 Pro launch chart, where Kimi's own runs used the Claude Code harness at max effort and competitor scores were 'taken as the maximum value reported in either their blog posts or the official leaderboard.'

Read across, the pattern is consistent: K3.1 Pro is at or near Opus 5.5 on agentic software engineering, clearly ahead of Astra on the two terminal-style benches Moonshot chose to lead with, and clearly behind Astra on science-flavored terminal work and Moonshot's own long-horizon test. Two numbers deserve suspicion in opposite directions. Moonshot's 66.4 for Opus is higher than the official board's 64.85 because the chart takes the best figure Anthropic ever published, which flatters Opus and makes Kimi's 'narrowly behind' narrower than it looks. And Kimi's 65.7 came through Anthropic's own harness, which is the strongest available and not what most Kimi users will run it in. Artificial Analysis's independent 67 is the number that says the coding claim is probably real.

Overall index, speed, cost per task, and what they don't know

Two honest gaps in this table: Artificial Analysis's launch note didn't include Astra's index score or Opus's cost per task, and we won't fill those from memory; they'll be added when read directly from the board. The hallucination row is Moonshot's own test and ranks Astra best, Kimi second, and Opus a distant third, which is the kind of unflattering-to-itself ordering that makes the rest of the chart a little more credible.

The skeptics, with the numbers

• 'Wait for the benches.' Correct on the facts: as of October 23, K3.1 Pro is on exactly one independent board. LiveBench, Arena, tbench.ai, SWE-bench Pro, and Vals have not scored it. Moonshot's chart also omits Gemini 4 Argon, which leads Arena at 1,517.
• 'The harness did the work.' Partly. Kimi's headline coding numbers came through Claude Code, Anthropic's harness, which is the best agentic scaffold available. That's a legitimate way to show a model's ceiling and a poor guide to what you'll get in Kimi Code or a generic agent loop. Artificial Analysis's 67 on its own harness narrows this objection.
• 'Open weights, but you can't run it.' True. K3 was about 1.4 TB with a sixteen-GPU validated minimum, and K3.1 Pro has the same architecture. 'Open' here means auditable, rentable from any host, and fine-tunable by labs; it does not mean local.
• 'The weights aren't out.' Also true. 'By Nov 4' is a promise, the license is unstated, and K3's license carried attribution and revenue clauses. Price your plans on the API until the repo exists.
• '$2/$10 is the real story.' The strongest point in Kimi's favor. At max effort, agent runs are output-heavy; Kimi's output rate is half of Opus's and a fifth of Astra's, with no long-context multiplier. For a team running thousands of agent tasks a day, that gap is the budget.

Which one, for what

• Autonomous coding at volume, cost matters: Kimi K3.1 Pro, via the API, OpenRouter, or Vercel's gateway, with a week of your own evals before you commit. It is the first open-weight model that belongs in this sentence.
• Autonomous coding, ceiling matters more than cost: Claude Opus 5.5 in Claude Code, still #1 on the official Terminal-Bench board and in our code-editors ranking.
• Lowest hallucination on retrieval-heavy work, and Codex: GPT-6 Astra, which led Moonshot's own hallucination test and the science terminal bench, at five times Kimi's price.
• Million-token documents without a surcharge: Kimi or Opus. Astra's 272K threshold makes long prompts the most expensive thing in this table.
• Fine-tuning, auditing, or hosting your own: Kimi, after November 4, if the license allows your use and you have the GPUs.
• A chat subscription for everyday work: none of the above specifically; see the chatbots ranking, where the app around the model matters more than three index points.

Our read

The comparison people want is 'is Kimi as good as Opus and Astra,' and the comparison that matters is 'is Kimi good enough at a fifth of the price.' On the first, the evidence is Moonshot's chart plus one independent index: close to Opus, ahead of Astra on agentic coding, behind on science and long-horizon work, three index points short overall. On the second, the answer already looks like yes for a large class of agent workloads, and that is what will move usage, the same way DeepSeek's pricing did in early 2025. What would change our read: the weights arriving with a clean license on November 4, which would make Kimi the obvious base for every US inference host; or the independent boards landing it well below Moonshot's chart, which would turn this launch into K3's story again. We'll update both this table and the explainer as each board posts. Related: Muse Wonder vs Opus 5.5 vs Astra, the previous entry in this series, and the token counter, where K3.1 Pro's price now sits beside the others so you can cost a prompt across all three.